Athens University Economics and Business (AUEB)
A modelling approach for correlated binary outcomes
Abstract
dc:descriptionThe scope of this Thesis is to contribute in the area of modelling the correlation structure of clustered binary data. The model developed is based on a joint likelihood representation where within cluster correlation has been incorporated as a function of a cluster specific parameter and the cluster's marginal probabilities, with the parameters of interest being the correlation parameter together with the regression coefficients corresponding to marginal expectations of the binary responses. A constructive technique, proposed by Oman and Zucker, is used in order to define binary variables with given marginal probabilities and some simple parametric pairwise correlation structures derived from the variables' joint probabilities. In particular, the parameter related to the correlation expresses the relative position of pairwise joint probability between the independent case and the one of maximum correlation. In addition the modelling approach is a joint likelihood method using the Bahadur's representation which is based on the decomposition of joint distribution into the joint distribution under independence, and a correction factor that incorporates within cluster correlation. In the current approach the 2-order approximation of joint distribution will be used, assuming correlations higher than pairwise to be equal to 0. For the estimation of model’s parameters, Marcov Chain Monte Carlo Bayesian methods are used due to being difficult to be analytically tractable. Various modelling schemas of the correlation parameter are examined with the most efficient ones being the simple model with a common correlation parameter for all clusters and the model where the cluster correlation parameter is modelled as a function of a linear combination of covariates. Further models have been examined where marginal probabilities are modelled through observation or cluster random effects or correlation parameter has a cluster specific random effect, but convergence issues were confronted. The models' performance is examined in simulated and real datasets.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Athanasopoulou, Ioanna
- Αθανασοπούλου, Ιωάννα
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/47430